Early Recognition Approaches

The earliest approaches concerned recognition within small populations, with the volume of data limited largely by the computational resources available then. As illustrated by Fig. 4.1, many sought to derive a human silhouette from an image and, as common in pattern recognition, and then to derive a description which can be associated with the identity of the subject. This gives a ‘statistical’ approach to automatic gait recognition wherein the image sequence is described as a whole, and neither by a model- nor by a motion-based approach, but one which describes the motion content. Given that techniques to separate a moving object from its background have been developed for many years (especially in surveillance) there existed a selection of approaches to derive the silhouette of the moving subject, an example silhouette extraction from the image of Fig. 4.1(a) is shown in Fig. 4.1(b). What was then needed was approaches to process the sequence of silhouettes so as to derive a gait signature — a set of numbers to represent a subject’s identity. A selection of signatures is shown in Fig. 4.1(c) which are displayed in a 3-dimensional space for visualization purposes only (the signature usually has more dimensions than this). These signatures represent ten of walking subjects with four instances of each subject. In this case the signatures derived from the four instances mostly cluster well, representing a good recognition rate. Accordingly, the target of the early approaches was to process video data of the form of Fig. 4.1(a) to obtain silhouettes of the form of Fig. 4.1(b) so as to derive signatures for recognition performance analysis.